• Z
    zhousa01

    发现有tf版本shuffler+centerNet可以在VOC数据集上收敛。我用CenterMaserPro无法复原。
    后端都一样:shufflenet_v2_x1_0
    Head前部分(pytorch)
    num_output = self.cfg.MODEL.CENTERNET.NUM_CLASSES
    self.cls_head = nn.Sequential(
    nn.Conv2d(256, self.cfg.MODEL.CENTERNET.HEAD_CONV,
    kernel_size=3, padding=1, bias=True),
    nn.ReLU(inplace=True),
    nn.Conv2d(self.cfg.MODEL.CENTERNET.HEAD_CONV, num_output,
    kernel_size=1, stride=1, padding=0))
    self.wh_head = SingleHead(256, 2)
    self.reg_head = SingleHead(256, 2)
    Head前部分(TF)
    with tf.variable_scope('feature_map_fuse'):
    deconv1 = deconv_bn_relu(out_32, cfg.feature_channels)
    out_16 = conv_bn_relu(out_16, cfg.feature_channels, 1)
    fuse1 = deconv1 + out_16

            deconv2 = deconv_bn_relu(fuse1, cfg.feature_channels)
            out_8 = conv_bn_relu(out_8, cfg.feature_channels, 1)
            fuse2 = out_8 + deconv2
    
            deconv3 = deconv_bn_relu(fuse2, cfg.feature_channels)
            out_4 = conv_bn_relu(out_4, cfg.feature_channels, 1)
            fuse3 = out_4 + deconv3

    发布在 社区求助区(SOS!!) 阅读更多

与 ManaAI社区 的连接断开,我们正在尝试重连,请耐心等待